A speech synthesis system that converts digital character data has been proposed. In order to convert English text or the like into synthetic speech, language analysis is required to extract information such as pronunciations and the like. The language analysis normally uses a dictionary for language analysis.
However, this dictionary normally includes several ten to thousands of lexical items required to analyze a language such as Japanese or the like, and has a large data size. Hence, when such dictionary is stored on a computer, a memory of several M bytes is required, thus disturbing efficient use of the memory.
On the other hand, the use frequencies of words (so-called trendy words), which reflect ever-changing social situation, become temporarily high, and such words must be registered in the dictionary with high necessity. However, the use frequencies of such words become very low after a given period. If such words are semi-permanently registered in the dictionary, the number of wasteful data increases, and such data cause analysis errors, resulting in analysis precision drop.